← All posts

For boards

AI is helping develop AI. What should boards do with that information?

OpenAI and METR report AI assistance in AI research. What does the evidence mean when a Norwegian company signs a long contract?

A hand checks a contract beside folders labelled OpenAI and Anthropic.

Imagine a Norwegian company about to sign a three-year agreement for an AI service. The product works well enough today, but the contract makes it expensive to change models or leave. If capabilities and prices change faster than expected, the company may be locked into an outdated service. If it waits for perfect certainty, it may delay useful improvements for customers and employees.

The board's decision is not whether to predict the arrival of superintelligence. It is how much flexibility the company needs when making a material commitment to a technology that may change quickly. Recent research offers a reason to revisit that decision, along with reasons to be careful about what the numbers mean.

What the laboratories report

On 18 November 2026, OpenAI published internal measurements of how its researchers use coding agents in AI development. It estimated that, from March to October, the share of delegated research tasks completed without human intervention rose from 23% to 46% for tasks it estimated would take a person two to four hours.

That is not the share of all research completed by AI. It concerns work researchers chose to delegate, within OpenAI's own environment. OpenAI says that the types of tasks delegated may have changed over time, so the increase can reflect both improving capability and different choices about which tasks to give agents. Its estimates are company-reported, not an independent measurement of ordinary Norwegian work.

The results also varied by task. For the period from August to October, OpenAI estimated a 79% success rate without intervention for monitoring training runs for problems, compared with 25% for writing inference and serving code. Researchers still described close human involvement and failures requiring correction. These figures should not be transferred to a company's customer service, accounting or software development without testing the actual work.

A separate assessment published on 5 November by the independent research organisation METR examined AI research at Anthropic. METR concluded that progress on well-specified tasks was likely faster than its historical rate and judged AI assistance likely contributed, but stressed substantial uncertainty. Anthropic could redact non-public information, and METR says much of the evidence it found most informative was withheld from the public. Readers therefore cannot independently inspect all the evidence behind its conclusion.

The studies examine different organisations, periods and measurements. They do not establish a universal growth rate, prove a self-improving system is operating without people, or tell a Norwegian board when a particular capability will arrive. OpenAI also says its reported trends may not predict future progress. The signal is narrower: AI is already contributing to some of the work of developing AI, while human judgement and practical constraints remain important.

The contract decision

Suppose the company is an engineering firm buying an assistant to search maintenance documentation. It expects the service to reduce time spent finding information, while a qualified employee remains responsible for the answer. The company and numbers below are hypothetical.

The provider offers a three-year commitment. Before approving it, management should ask for answers to five questions:

  • What can change? Can the provider replace the underlying model, alter its capabilities or change where information is processed? What notice and testing are available?
  • What is the exit cost? Identify minimum commitments, fees, notice periods and the cost of moving the service. Ask for the answers in the contract, not only in a sales presentation.
  • What can the company take with it? Establish how to export source documents, configurations, evaluation cases and useful records in workable formats. Test a small export before relying on it.
  • How will changes be tested? Agree who checks Norwegian-language quality, errors on important documents, access controls and response time before a new model replaces the old one. A new version can improve one task and worsen another.
  • What happens during an interruption? Decide which work can continue manually, how the team reaches the documents and who authorises a fallback. Check the cost of keeping that option usable.

The strongest protections may cost more or be unavailable from a particular supplier. A smaller business may reasonably accept some dependence in exchange for a lower price or better support. The important step is to identify what the company is buying and what it would take to change course.

Put a value on flexibility

Suppose the fixed commitment is 600,000 kroner over three years. A more flexible offer costs 60,000 kroner more but permits a review and exit after the first year under agreed terms. Those figures are illustrative, not market prices. The extra 60,000 kroner buys an option, not a guaranteed saving.

Management should compare that premium with the plausible costs of staying in an unsuitable service, migrating, retesting and maintaining two solutions. It should also include the value of getting a useful tool into employees' hands now. The flexible offer is not automatically best: the extra fee may exceed the benefit, the exit may still be difficult, or a longer commitment may fund needed supplier support.

Record the assumptions before signing. At the review date, check whether the tool has actually improved the selected task, whether customer and employee requirements have changed, and whether another option is both better and safe enough to adopt. A benchmark alone is not the trigger to migrate.

Keep the board decision proportional

Management owns supplier due diligence, security, a representative test set, the contract negotiation and the fallback. The board should ask which assumptions could make the investment uneconomic, the maximum commitment before review, and whether the company can leave without losing essential information or stopping customer work. For a small purchase, that discussion may belong with management rather than the full board.

An earlier guide to board scenarios considers several possible AI futures. This decision is more immediate: before accepting a long commitment, determine what can be reviewed, tested and changed. The supplier exit guide helps turn that right into a workable plan. Contact Jonas to discuss an AI investment or supplier decision.

Sources checked on 20 November 2026. The company, contract offers and amounts are hypothetical. OpenAI's figures describe its own research environment; METR's assessment contains important redactions and uncertainty. The recommendations are the author's analysis. The illustration is AI-generated.

Put ideas into practice

Work with me

Explore what AI and cybersecurity mean for you and your business. Find options for a confidential session, ongoing advice, or practical collaboration.

Explore the options